How to select the models to use?

[haimat]

I would like to use docTR-Labeler to annoate our custom dataset. Is there a way to specify which detection and recognition models the application should use for the AUTO ANNOTATE button?

[felixdittrich92]

Not yet with other models as fast_ there seems to be an issue I have to fix before.

The easiest would be to write a quick script on your own:

For example if you download reference datasets

and inside the detection_task/train folder rename the labels.json to pre_annotations.json and open the detection_task/train/images folder with the tool you will see that the annotations are imported correctly.

So everything you would need to do is to pre-label your data with docTR and save it in the correct format (images folder and the corresponding pre_annotations.json) then you can open it with the tool and correct.

NOTE: Also already complete correct pre annotated images needs to be saved inside the tool otherwise the merged file after all images are done at the end is incomplete !

[haimat]

Thanks, that did the trick for me to debug this issue.
For some reason the command line has loaded doctr-labeler from the user's pip installation folder.
I had to restart the venv in order to run the version I installed there.
Now the models from command line are being loaded correctly 👍

[felixdittrich92]

Ah ok 👍 Does this also fix the issue now ? :)

[haimat]

Yes, thanks, looks better now.
Next I will annotate our dataset and fine-tune one of your models to test 👍

[felixdittrich92]

Nice glad to hear :) 👍

[haimat]

Thanks for the quick help, I appreciate that very much!